From: AIPS 2000 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All rights reserved.
Mixed,-Initiative
Resource Manage inent:
Marcel
A. Becker
+u,~t
Stephen
"[’h,’ lhA.,,~tic.-. Iusritul.,’
Carnegie M~,ll, m l’uiv,,rsity
5000 Forbes Av,.nu+, Pit t,d>urgh.
rob81 sf.sh,.s omu.edu
(412) 268-,’1811
Abstract
!n Ill[,. palwr, we drscrib,. Ihe Barr’<l..lll:,,’a£of,
:’..,.cheduling Leo[ drvt:lopt-d for day-to-day allocation m’d m",Jatagemrtlt of airlift and tmhker
l’t’,’.,.OllrC,’.’..,
ilt the [’SAI" .’kit" Mobility
{’<)lrllll~.~lld
(AMt’j. The syat, en.i utilizes ml illcrenmnt+~l arid
configura[~i,z con~trailit-l,a.-,cd ~carch frameworkto
pi’O~,iduar.;ttlgc of &UtOl|l,tt.t:d ;.u|d ~t:mi-,J.lit,)tltated
-whrduhng c;l,pabilities,
including generating an
initial solution It) the fleet assignmenl problrm,
selL.clive re-optimize!ion el resource allocations
uo i:lcorporare new higher priority missions while
nlininlizing sohlt.ion C]lallge,
mergingo[" previously
planned mission.-, t.o reduce non-productive flying t,ime, attd generation and synchronization of
tanker missions to satisfy air refueling requiremerits. In situations where all mission requirements calulot b(! incL. the system can generate
mlcl compareel.It ernalive constraint relaxation eel ions. The currrnt version of Barrel Allocator will
t
go illto operational use el. ¯AMC
as a module of
(-~[{Please 2.0 of AM. s Consolidated Air Mobility
Planning System (CAMPS) in early 2000.
Introduction
l’3li,’ient
allocation of aircraft and crews to transportation missions is an important priority el. the USAFAir
MobiLity (’omnmnd (AMC), where airlift
drmand must
increasingly be met. with less capacity attd at lower cost.
The A MC resource management problem presents
several interesting challenges:
¯ problem scale - Over a typical short-term (e.g., 2,
week) scheduling horizon, several thousand air ulissions are flown world-wide, utilizing
several hundred
air<’raft and active-duty air crews. In titles of crisis.
these numbers can increase substantially,
and additionally inwflve both reserve units and comntercial
aircraft.
¯ probletu c’.omplexity - Resources must be allocated to
missions in a way that minimizes non-productive flying time. attends to mission priorities,
attd maximizes
*Authors listed in Mphabetic order
(:opyright (~) 2000, American Association for Artifici,d
tcUigence (www.aaai.org). All rights reserved.
32
AIPS-2000
The AMCBarrel Allocator
F. Smith"
PA 15213
the nunlb,:r or" support,.d missions,whih. at i.lt,, sam,.
time rnsuring that decisions are fra.-+ibh’ with r,"Sl,,’<’t
to aircraft availability at,d operating ,’l,;.u’aetcri..,ti,’s.
crew ,t,ity clay limits, required mission :,x,cul ion windows.airport capacity and lauclitl..m.t.ittte r,’sl ri,’l ions.
and other mission ,:ot~str:.tit~ts. In ,’ase.-: wht.rv all ,’oustraints cannot be satisfied, sol[re may br select iv+.ly
rclax<..d +mdtradeofEs thus[ be LlJ.ad,: bctwv,.nt ah,,rua! iv<, opt ions.
¯ solution continuity- Like most pra,’l i,’al <lot[rains. rt.sourc,, allovation is sil.ual ed in a cent iutt,m..,ly ,.x,’,’uting r.llvironxnent.I’:a.<’h tim,. a rt’,’,~:,lUL’C¢’;L..-Sigliill,’ull iS
+
chaltged,
llQW orders Ill[IS[
bC’ r,’-cOIllnllllliC;Ll+’<l
IO th<
affected wings and r++-assimilat,,d into lo<’ally plalm,+d
m’tivitics,l-I,-nc,,, il ix ]llll)r+t’t;lllt
tt:, Itl;Itl:’lg,’
:|ll, I ztninin,izo solution ,.’]tange as now ,,issions ,r,. in,’,,rporated into the scheduh+ over time.
¯ interacting planning and scheduling prol31,:nts- I:ll’t.ctire allocation of resour<:es requires a tight interl+lay
of planning and schedt,ling capabilities.
By ,h.fm.th,
missionsart. i)lanncd a.s round trip el)[.rations and
eacli assignedaircraft returt,s to its homebast, upc,n
comp/etio||. [lowever, lift capacity can ,.~fteu I., incroas<,d by contbinlitlg two or moreLnission.s :tttcl "’r,.’cycling" l.ho satlle aircraft front one ntissi,.m I,, uh,,
nrxt. Aht’rnatively. a l)lanno, l airlift missi,.,u nlmy r,,quireairrefueling,
hl thiscase.tatuker
capa,’it.y
mtust
b,+~sourced, and a slipporting
t, al’lker
missioJt must I)i,
generated
and sym’l’m:mized
in tim+,.
Duo t.o the time pressure of decision-m+~king
and tit,.
lack of autotnat~’d scheduling tools, Ihr AMC."’Bartel Masters" responsible for tnakittg allocal.ion decision~
typically make allocation decisions in a litttilrd,
myopic
fashion and routinely miss oplmrt unities to opt imize r,:source usage.
In this paprr we descril~+, 13arr,,l All,~,’at,Jr.
a t,~ol
for generating and evaluating such optimizat ion ol>portunities.
Dev<,loped through applicalion o[" thr ()zon,
scheduling framework (Smith ct ,l. 1990; Ik’<’k,.r 1998).
Barrel Allocator utilizes incremmlt el, con.,,l.raiul-l.mso,
I
s<’heduling techniques to allow integration of n,,w tnissions and response to changing requirements arid availability,
while minirnizing disruption to most previous
From: AIPS 2000 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All rights
ally reserved.
specify which wing they
assignments. In situations where all constraints cannot be satisfied, the Barrel Allocator’s search model
promotes selective constraint relaxation to find an acceptable solution..Mission scheduling and resource allocation capabilities can be invoked in automated or
semi-automated modes; in the latter case, the system
generates and compares different options that might be
taken by the user.
Experimental results with Barrel Allocator, obtained using historical data extractcd from the current
ADANS
airlift planning system indicate the potential
for substantial improvement in resource usage, through
better optimization of air wing assignments, selective
combinationof missions to efficiently "r~,cycle" aircraft,
and more effective integration of tanker and airlift missions. Following positive review by AMCpersonnel, a
version of Barrel Allocator has been delivered to the
Tanker Airlift
CommandCenter (TACC) at AMCfor
extended user review and testing. Current plans call for
Barrel Allocator to go into operational use withill the
"I\a.CC early next year as part of release 2.0 of AMC,’s
Consolidated Air Mobility Planning System (CAMPS).
The remainder of this paper is organized as follows.
Wefirst briefly summarizethe Barrel Master allocation
problem and current practice at AMC.’this is followed
by an overview of the functional capabilities provided
by the Barrel Allocator tool. In the next two sections,
we discuss the underlying representations and search
procedures that provide the system’s technical basis.
Wethen provide some details on the status of the hnplementation and technology transition effort. Finally,
we discuss some planned extensio~ls.
The AMCBarrel Master Allocation
Problem
The AMC"Barrel Master" (or Barrel for short) is
charge of resource allocation and resourcc management
for the USAFAir Mobility Command. The different
planning offices at AMCsubmit resource allocation rr~
quests to the Barrel in the form of missions or lnission requests. Thcse missions represent, for example.
requests to move cargo and/or personnel, or requcsts
to reserve resources for a numberof training activities
and exercises. Althoughdifferent types of missions create different types of resource requirements., all planned
missions specify a particular type of aircraft to be used,
an itinerary, a priority, a preferred air force unit or wing
to fly the mission, and a time period, represented as a
set of dates, in which the mission should be executed.
Each Barrel manages particular sets of aircraft and
corresponding crews. These ~ts are defined by aircraft
type and by the geographic locations where the aircraft
are stationed. A set of aircraft, of the same type stationed at a particular air force base constitute a wiT, g.
For example, McGuire Air Force base has a wing of
C141s and a wing of KC135s. Currently, one Barrel is
responsible for all west coast C141s and another Barrcl
is rcsponsible for all east coast C141s. Planners gener-
would prefer to fly a givcn
mission, and when possible, the Barrel will allocate a
plane from this planned wing.
The mission itinerary is the sequcnce of stops or airports the aircraft should visit during the exe,:ution of
the mission. Werefer to the flight between two successive stops as a mission leg or just a leg. Eachleg has an
origin airport, the Point of Embarkation (POE), and
destination airport, the Point of Debarkation (POD).
Each leg is followed (or preceded) by a certain ground
time. During the period the aircraft is on the ground,
a nmnber of activities, or g~wmdevents, can occur: for
example, loading and oflloading of cargo, refueling, crew
rest, crew change. The time period specified ira tile ntission request should be at least as large as the time required by the aircraft, to fly between all intt’rmediatc
stops in the itinerary plus the required ground time at
each airport. The Barrel will try to assign one air,:raft
that is available during this entire period. The earliest
date the mission can start is called the Available to Load
Date (A LD) and the latest date the mission should finish is called the Latest Arrived Date (LAD).The length
of this intcrval should be at least as large as t lw total
duration of the mission.
Aircraft availability is defined for eaclt wingon a daily
basis. Each wing has a total number of aircraft of a
particular type. Considering that sonic planes ar~, undergoing maintcnance and the wing has some need for
training and local missions, the wing will make a subsct of all its planes available to AM(’. missions. Each
day, each wing will provide a certain number of conhuct aircraft that can be allocated by tire Barrel. The
remainingaircraft, designated as fenced aircraft, arc: reserved for local wing use and are beyo,ld thejurisdictio,
of the Barrel.
In her/his daily activities,
the Barrel Master currently managesresource availability using a ~’omT~itmerit matrix. This matrix tracks available aircraft c~,pacity of different wings over time and records those
missions already allocated. As ucw missions are r~
ceived from various planning offices, the Barrel consults
this matrix and tries to allocate r~.sources which satisfy
mission requirements. If all requirements can be satisfied, s/he makes the aircraft assignment and communicates mission commitmcntback to the planner. In those
cases where there are insulficient resources available to
support a particular set of missions, s/he will consider
more disrupt]vc allocation alternatives. For example,
s/he will considcr using resources already allocated to
lower priority missions or will cot,sider using resources
provided by a different wing. Once one or more acceptable options are found, the Barrel communicates these
possibilities back to the relevant planner and a solution
that wouldbest satist~" all sides is negotiated.
Thc AMCBarrel Master problem is similar to tl,e
problem known in the ORlitcrature
as the Fh’¢l Assignment Problem: Given a schedule of flights defining
the departure and arrival times for each fly leg, the Fleet
Assignment Problem is the problem of deciding which
From:
AIPS equipnmnt,
2000 Proceedings.
Copyright
© 2000,
(www.aaai.org).
flight
or fleet,
should
be AAAI
assigned
to each All rights
In reserved.
all cases,
flight segment (Barnhart ,’l el. 1998; L.W. et el. 1996:
R.ushnleier and Knotogiorgis 1997). The AMCBarrel
Master probhun addressed by the Barrel Allocator is
a dynanai,’, string-bas,.d version of this problem. For
commercial airlines, the objective is to maximize revenues minus operating costs. The Barrel Master tries
to maximizell,e total sum of priorities: s/he will try to
assign the maxinmm
re,tuber of high priority missions,
a, td would only consider assigning lower priority missions after all higher priority ones have been assigned.
Tl’aditional OR-basedsolutions to the fleet assignnmmproblemassign fleets to inclividual flight segments.
The Barrel. alternatiw’ly, is concerned with assigning
fleets to a sequence of segmentsor slrings. A slring is a
SeCluenceof conneclecl flight. Sogl,,el,lS
that begins and
ends at possibly d ilferent maintenance stations (B;rrnhart el el. 1998). An AMCmission itinerary is typically planned aa a string that starts and ends at the
seine location (i.e.. a rot,nd trip). Strings that start
awl end at the same station are usually referred as aircroft rotations. If possible, the Barrel would consider,
anti so]not imes even I)refer. using t he same total.ion for
more that, one mission. The di[ficulty in combiuing
missions is in identifying the opportunities for potential ,’ombim~tions amongthousands of missions in the
database.
The hlarrel Ma.ster in charge of refueling resources,
the "l’ankt, r Barrel, in addition to allocating t.ankors to
planned air refueling missions, is also responsible for
linking air vefuelin 9 events with regular airlift missions.
Anair refiwling event is a request for a refiteling mission
that is generated each lime a planned airlift mission requires air refueling. In the data repositories currently
available, there is no explicit linkage betweenair refueling eve,,ts and the missions Ihcy ~Lre supposed to serve.
Thus. the Tanker Barrel currenl,ly has t.o perform this
linkage by manually searching the database for airlift
missions that matches the location and time of air re,fueling events.
The allocation process described in previous paragr;~phs is currently mostly manual. The Barr~q Master uses ~t system lhat provides an electronic commitmeat matrix at,d is linked to current aircraft availabilil.y and mission data. However, resource assignment is
performed one mission at. a time with little automation.
Capabilities fi)r identi~.’ing tnission combinationopportunities are quit~, limited and ~s just meutioned there
is no sysl.em support, for establishing linkages between
tanker and airlift rnissions. The Barrel Allocator system described below aims at automating some of these
tasks and enhancing t he decision-making capabilities of
t.he Barrel master, while still granting fifll control and
visibility over tl,e dc.cision makingprocess.
Functional
Capabilities
The Barrel Allocator provides three core sets of fuactionality to the AMCBarrel Master: n:souwe allocalion, mission combination, and air refueling linkage.
34
AIPS-2000
functionality can be utilized in a more
or less automated fashion, ranging from a fully manual mode where the system does lit.tie
more than decision bookkeeping, to a semi-automatic mode. where
the system generates alternative options and previews
their impa,zt, to a completely automatic mode, where
the system determines selects the best decisions based
oo user-specified preferences. In the paragraphs below.
we first summarizethese core functional capabilities.
In subsequent sections, we then discuss their tm’h,ical
basis.
Resource Allocation
In a typical modeof operation the prol)lem is one of
integrating sets of newly planned missions into an existing current global schedule. In this mode,tl,e ~chcd,I,r will attempt to assign aircraft and schedule new
missions without disrupting the current set of assignments. Any mission that cannot be integrated into the
schedule in this way (which implies that there is not
enough lift capacity to accomplish the mission without changing exisling resource assignments) is flagged
as nnassignable and will require subsequent ,s,.r attention.
The assignment of wings to new missions can be performed in manual or automatic mode. The userselects
some set of unassigned missions, the alternative wings
to be considered fortl,ose missions, and the system will
,:,)inlmte all feasible allocations. In rnanual mode.t, he
application displays all ft, asible solutions and the user
can select the preferred one. In automatic nmde the
system selects the best allocation based on some vser
defined preference. (for example., try to us,. tl,e wingthe
minimizes ow’rall mission time weighted by priority).
The allocation process produces an assignment of (notional) aircraft froLn specific wings to each mission, and
an assigmnentof flight times to each mission leg. If t l,.
locations corresponding t,(,origin and destination of the
mission are different from the location of the wing providing the plane, positioning and d,.-i)ositioning Ilights
will be added to the itinerary.
Currently, the following constraints are taken into account and enforced in consl.ructing a schedule for all
current missions:
Wing capacity coastraints
- Assignn,enL of ,nissions to wingsdoes not exceed the m,tnber of<’ont ract
aircraft available at each wing.
Mission time requirements
- Missions must be
scheduled within time windows designated by ALD
and LADco,straints.
Enfi)rcemcnt of required ground time - The allocated time ;~:counts h)r aircraft onloa~t, otfload and
minimumtime-on-ground constraints, each specified
as a function of aircraft type. More generally, it is
possible to specify a range of different flight preparation activities and tiv,e constraints.
Flight duration constraints - flight duratio,, cat,
From: AIPS 2000 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All rights reserved.
be specified as an input data or is computed using
great circle route
Aircraft
range constraintswhich are enforced
when determining specific wing assignments (and
hence when creating positioning/de-positioning
flights)
Crew duty day constraintsDepending on designated crew type - basic or augmented - crew rest
is inscrted at appropriate intermediate points of the
mission to enforce crew duty day.
As suggested above, it is not always possible to satisiS" mission time constraints with existing available lift.
capacity. To support resohltion of such situations, the
system gives the user the ability to analyze and compare various constraint relaxation options. Specifically,
for any given set. of uuassignablc missions, t he user can
choose to:
Allow bmnping of lower priority
missions Every mission has a pre-defined priority. If there
is no aircraft available to support a high priority
missiom one ahcrnative is to pre-empt lower priority
missions already in the schedule. Any pre-empted
missions arc then rescheduled in succession and they
may. in turn. pre-empt missions of lower priority
still. At quiescen,:e, any lower priority missions that
cannot be be re-inserted into the schedule within
its constraints are added to the unassiguable list. h
mission locking mechanism is provided to allow thc
user to avoid bumping any specific, lower priority
mission. A fivem, interval,
a period of time in
whi,’h no mission can be bumped, is also enforced.
The freeze interval is required to avoid schedule
turbuh’m’e close to execution.
Over-allocate - Since the number of contract aircraft
is usually smaller than the total number of possessed
aircraft., the user mayalternatively choose Io go over
the published contract level of a given wing. This
happens with a fair anmuntof frequen,’y. It typically
reflects extra knowledgethat. the barrel master may
have about wing assets or agreement on the part. of
the wingto use fen,red aircraft.
Delay the mission- The user may consider the option of delaying the current mission until necessary
resources are available. If delay seems like a potentially viable alternative then this information ,:an be
suggested to the mission planner. Mechanisms for
limiting the amount of delay acceptable and combinations of delay and bumping, and delay and overallocation are also supported.
Use alternative
MDS-Similarly,
it might be possible to accommodatethe mission if an alternative
airframe type can be utilized.
Any of these options can be invoked by the user in
"what-if" mode; a general "undo" capability allows the
retraction of any sequence of scheduling actions that
have been issued by the user, and thus provide a basis
for exploring alternatives. Alternatively, the user can
ask the system to generate all options and compare alternatives. This range of scheduling modes provides a
continuum of scheduling actions that are progressively
more disruptive in the changes that can be made to the
current schedule. There is no formal metrics to compare
alternative relaxation options. Each option will havc its
own disruption metric (e.g., anmunt of resource overallocation, numberof hours late, numberof lower priority
missions bumped). The user is currently responsible for
evaluating and selecting the best alternative based on
his/her ownsubjective criteria.
Mission
Combination
Missions are planned by default aa round trips from a
particular homebase. If the origin and/or destination
of the mission does not coincide with this base, positioning and depositioning flights segments ar~ added to
the mission itinerary. To improve resource utilization
and increase aircraft, availability over time, the Barrel
Allocator provides the capability of exploiting mission
coml)iuations: it will look for opl)ortuuities to link two
missions su,:h that after the end of one mission, the airplane will be "recycled" and redirected to ~upport another mission instead of depositioniug back to its home
base.
The mission combination capability can be used as
an alternative allocation policy to allo,’ate unassignable
missions or as a compression mechanismto increase the
number of airplanes available over a certain period of
time. The user can select or filter the potential n,erge
candidates using three different parameters:
Maximum layover time - the maximum time delay
that can be tolerated between offload (end) of Ol,e
mission and onloaxl (beginning) of second mission
Maxinmln distance-the
maximum distance
that
can be tolerated betweeu location of first missions’s
of Iload and second’s onload
Percentage decrease in overall flying time - the
reduction obtained by combining two misbior, s into
one aircraft, rotation comparedto flying both original
round trips.
Air Refueling
Linkage
’lb support the "additional responsibilities of the Tanker
Barrel, capabilities are also provided for generating aud
assigning tanker missions to airlift, missions that require
air refueling support. A set of refueling events is accepted as input, aaad tanker assignments can be generated either interactively or automatically, lu atttomatic
mode, the scheduler first attempts to link tanker missions that are already included in the schedule (e.g.,
training missions); tanker wing assignments are determined and tanker missions are created for any remaining unsupported airlift missions.
In interactive mode, a map-based display is used to
indicate candidate refueling tracks that are already covered by tanker missions in the temporal iuterval re-
From: AIPS 2000 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All rights reserved.
quired by a given airlift mission. The user call selc.ct
one of these highlighted r~fueling tracks (potent.tally
changing the originally phmnedrefueling location) or
select the planued track (which will result ill t.he creation of a newtanker mission if one is not already ill I.he
schedule). In the ca.se of the selection ol’a pre-existing
lanker mission, the system will also present opportunities to support multiple air refueling events with the
same tanker mission (provided that fuel reqtfirements
mtd tanker fnel capacity ,-onsl raints are sat isfied).
Problem Formulation and
Representation
As indicat.ed at the outset, the Barrel Allocator has
bee. developed through use of the modeling primitives
and class library defined in the Ozonescheduling framework. I)evelopment of an application system in Ozone
involves three basic tasks:
I. Ideal.tidying l.ho outological entities in the domainrelevant 1.o the application donmin.
2. I)efining and &’veloping a domain model, by mapping the problem spe,’ific oat.elegy onto the Ozotw
scheduling ontology,
3. Instantiating and/or ext.ending Ozone probleni solving templates to address tile set of relevant constraints.
In this section we describe the Barrel Allocator’s domain mo&l. In the next, we consider its problem solving procedures.
Ozone domabJmodels are defined in terms of live hasic entities (Smith and Becker 1997): demands, actit:itics, products, resoutr.’es, and ,onstmints. The mission
requests sent. by the planners to the Barrel correspond
to the demands. They are used I;o t.ransla~e the exter,al requirements specified by the user into thn internal constraint model used by the scheduler. The
bmsi,: attributes of the mission request art’: a pair of
dates establishing the earliest date the mission should
start and the latest date it should finish; the type of
aircraft and preferred wing providing it; /lie missioxl
priority (which imposes a partial order on the set of
missions); the itinerary or set of flight segment.s the
mission should fly; and the mission type. The mission
type is the product (i.e., service) of interPst. Satisfaction of a mission request, implies that a certain type of
service has been provided, andthe produ,:t is the atttity
t.hat represenl, s this service within Ozonemodels. Services are provided through the execution of activities.
The flightsegments,or legs, and groundevents
aretheprincipal
types
ofactivities
inthisdomain.
Ear’[~
typeof mission
imposes
special
constraints
on howthe
activitics
shouldbe performed.
Forexample,
refueling
missions
andfuelsupplymissions
require
special
typcs
of resources and temporal synchronization between sets
of activities. Twospecial types of flight segments are
the refueling
leg and the fuelsupplyleg.The refueling
legcorresponds
tot.hem’tivity
ofreceiving
fiwl.
36 AIPS-2000
~nd the fuel supply correspondst.o one (,f providi.g fu,:l.
Suchproblem-spc.cific knowledgeabout the (’oiler r;tinl.s
associated with different mission lylws is encoded into
the product entity and ,s,.d d,ring it,s(:-mii;~ti,,n
:rod
scheduling of nfission ~wtivit.ies.
A flight
segment
a,’tivity requires lbur dill~.r,:ut
types of resources: I,h++ aircraft, the l.wo airports
correspondillg t.o origin and destination of tl,. Ilight,
and tile air crew. Asso,’iat.ed with ea,.’h resour,’,, is
the notion of <’almeil.y: the alH<Jnlfl or quantify ,,f a
<’ertain t.ype of unit, that is availabl,, overI.inw. Airt’raft
availability is definedat. the t~ggregalofleel level. AII<,et
or wing is a pool of aircraft of the same t.yp,. I,:’ated
at a ccrl, ain geographiclocal.ion. Ea,’h ai,’,’,’aft it, a
wingrepresents o,e unit of capacity. ’l’lw tot.al nmi,I),’r
of aircraft I,Mongingt.o a given wing delim.s its tol;al
capacity. Since each plane can only tly on,. luissit,n
at a time, tile tot.al capacity [s the lllaxiliillitt
re,tuber
of mi~qionsflint ,:an 113’ simuhan,.,ou.qy tit any given
time. Thetotal ¢’;~pa¢.’ity ,,f the wingis divi, l,,,.I it,i,,
I.wo subsel.s: I.he contract capacity, relJrc..,enting tl,,’
set of planes available to support AM(’missi,ms, and
lhe fenced capacity, represenliltg the .,,el ,,1" Id;Ul,’s
reserved for local missions and ,’armor I,,., u.,,.d hy llw
Barrel in normalcircumstal,c~,s. Weonly ,’h,~,’k "airport,,
for airplan,’ compat.ibilityr,,stri,’tio..,..q.’Ul)l,,)rt fi,r air
crews is not curret,tly supp,~rted and will b,. in,’h,d,’d
in the next version of Ihe systen~.
Resource capacity is defined ov,.r capacity
intervals.
Ea~-h interval has start, and end tirnes.
and vah,es for total, contr;u’t, fen,’t.d, and et, rn,ntly
availabl,- capacity. Tit,? end time is u.lways larg,’r than
the start time and the nfiniml, msize of an interval is
one time unit. Time is rel)resenle, I conlinu,,usly Io
the granularity of lhe time unit. For ,’xanil)h., Ihe
Barrel Allocator uses minules as t.lw. tim,: gran,la.rity.
Therefore. the minimumdistance Iml.w~,en two tin,.
points is one minute.
The o,tput of the Barrel Allocatc, r is a s,..t of
resourceassignmentsor resourcereservations:
tuples of the form (Mission,Wing, start-time,
end-time),
rl’his rneans thai ntission Mission
is r,’serving one unit of capacity of resource Wing during
the time interval starling at start-time and finishing
at end-time. The duration in nfinutes of this interval
the dist.a,ce betweenstart and t,l,<l tim,.s, is ,.qua[
to the total mission ,lur~ttion. This duration inchld,.s
tile duration of all ti,: flight segments, pl,s required
time tbr all ground e.vents. As we will discuss in /lie
next section, different constraints mayI)e satisli~.d by
this assignment, dependingon the set of parant,A.,,rs selected by the user while act.ivating a particular pml’,l,,m
solving metl,od.
Problem Solving Methods
The planning and schedulin¢,eY procedures used to provide various fimctional ,’apabilit.ies within the Barr,.,l
Allocator are instantiations of ba.si," searct, methodtemplates available in the Ozoneschohtling framework, lit
From: AIPS 2000 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All AssignMissions{Unassigned_Missions.
rights reserved.
config}
While Unassigned_Missions is not empty
Do
/
\
¯
I
.../
Evalce~m
Figure 1: Basic search procedure tbr resource allocation
(assuming that mission being assigned requires C141
aircraft with earliest pickup at ll and latest delivery by
t2).
Ozone. search templates are only partially instant.iated,
producing t)arameterizable problem solving procedures.
These procedures can be dynamically config||red to provide different search behavior as the problemsolving situation dictates. In the subsections below we describe
these core procedures and how they art, composed to
produce user level functionality.
Resource Allocation
The central function supported by the Barrel Allocator
is that of assigning planned missions to w i ngs over ti me.
Figure t graphically depicts the general search template
used to make this assignment for a single mission.
The search procedure proceeds in 3 steps:
1. a set of candidate resources (wings) is generated,
2. for each candidate wing, a set of possible allocation
intervals is generatcd, and
3. each <,ing, allocation interval> pair is evaluated and the highest ranked candidate is selected.
As implied by Figure 1, a fidi instant[at[on
of
the AssignMission search procedt, re is obtained by
specifying three components: a search operator for
generating candidate resources (referred to generically as GenRe.sources), a search operator for generating candidate allocation intcrvals (generically called
Gt’)~lnter~ats). and an evaluation metric (Evalcrm#ion
for ranking ahernatives. By parameter[zing the procedure ¢o operate with different sets of operators anti
evaluation criteria, resource assignments can be generated and evaluated under a range of different constraint
relaxation assumptions.
In the most basic case, AssignMission
is configured
to search only for feasible assignments, i.e., <wing,
allocation interval> pairs that are consistent with
the time and resource requirements specified by the
mission and are also compatible with the assignments
Extract a mission Mfrom Unassigned_Missions:
If AssignMission(M, config)
Then Mark M as assigned
Else Mark M as unassigmlble:
EndWhile
End
Figure 2: Overall Mission Scheduling Procedure
oi" previously scheduled missions. This feasible configuration of AssignMission
is obtained by incorl)oralion of the triple < (~enit, questedRrs, GCnPea,.iblL.Ints,
E~’alMinFlyingTime>.
Here, C,t:nReqnestedRes
gclleral.es
candidate wings consistent with aircraft type reqttested
by the nfission. Likewise,(;e nF,.~.~ibt,:tnr, scans a candidate wing’s capacity profile for allocation intervals (I)
with at le~Lst one unit of available capacity (i.e.. an aircraft.), (2) with a duration greater thaal or equal to
time req|,ired to ac,rOml)lish the mission (a function of
mission itinerary, aircraft, speed, wing’s homebas,’ location, crew rest r,,quirements, etc.), and (3) with starl
al,d end times that satisfies the mission’s earliest onload time and latest off-load time constraints. (’anti[dates are differentiated on the basis of total flying tim,,
and tht’ candidate assignment that minimizes this inettic is selected.
By selectively substituting different se~trvh operators
and/or evaluation criteria. AssignMission can ah ernatively be used to fi,d assignnwnts under various relaxed
problem assumptions. For some types of constraints.
relaxation simply implies the consideration of a dift~rent discrete set of options. For example, substitttlio,,
of eenAIternatit, eilts for Ve~tllLqut.stedR: s resull.s in generation of assignments that consider types of air,:rat’~
other than the type requested by the mission planner.
For other classes of constraints, however, relaxation is
more continuous in nature, and in substituting a search
operator that assumes constraints can be relaxed, the
search must also be biased to promote their satisfactiun
to the extent possible.
By varying the operator used to generate ;dlocatio,
intervals anti the evaluation metrit" used to prioritize
candidate solutions, a :mmber of useful AssignMission
configurations are defined:
Delay - Incorporation of t he triple <Genn,~,,~,t,.du,,.
(]CnDelaylnts, ~’l,’tllMinTardincss>
yields an assignment procedure which assumes that mission deadlines can be relaxed if necessary. GcnD..layl,ts uses
the same mechanismused by GenFeasibtel,,ts but considers a larger portion of the candidate wing’s capacity profile, and El,’ul:’~fin’.l"ardiness
ensures that t.he
mission deadline will be relaxed to the nainimunl extent possible
In this configuration advantage can I)e takel, of the
belween Copyright
seardl operalor
;rod(www.aaai.org).
evaluation All rights reserved.
From: relationship
AIPS 2000 Proceedings.
© 2000, AAAI
criterion to effectively constrain the number ofcatldiclare solutions generated. For ex;~mple, hy s(:atming
forward ill time t.ilroligil
U. resource’s capacity profih,.
Ihe Ill’st, interval with available cap;~city foum.l will
I., tho ¢,u, ~ thai lnininLizes delay for that resource.
If this approach is tak~n only one interval need be
g+i,,raled for e+wl] ,’audidat.o resmw(’,.. In ,>l.h,~r C(JILfigural.ions (e.g.. I.lit, pre-,ml,tion case below), wh<.’r+,
there is Ito such donlinan<:+, condition for constraintug soluuion gent..rm ion. tlloro a,l-ho," heurisl.i<’ <.utolr.s
<’till l," used.
()ver-alloeatc: - The
triple
<(Tt ttRf,,lucst ~ ,iRes,
(;enO,,.rlnt+, I’+’l’al.W~,c)c. erl:.,.+:],.> <-]¢fine8 an a.,~sigillllplil
proce, hlre where ca.pacity constraints al’p
relaxable. G¢no,,,.,.+,t.+ scans file cap+wily I)rofile of
a cmldidai.e wing. I,ut, g<.nrral, es allo<.’iuion int.erv~l.-,
that exl<’ltd al)ov,: th<" wittg’s "’<:ontracted" level (i.e..
,lipping inlo ils Io,’ally n,s,.rved or "felwe, t’" pool
uf aircraft cai)acit,y ). Ecal.,lli,,o7 ert’s,l:lt, prOiliOt.e~S
s,.le,’t i, ,n of t.h~’ geueralt,d alloc;d.iun inl.erval ll,;,i
lllinintizt.s the I<’vel of nv,,r-atlo,’ation.
lit this ,’its<:. ituL×inlal inlervals at dill’<’r<’nt
over-idlocat ion ,:an I,e ~’ffi,’i,’ntly
generated
+,;i.r scanof the l’OSt)urce’~,’;qiacit.y profi[<:,
sequeul.ly prllned to tnininiize the teinporal
over-allocai ion.
10:vols ¢,f
via a linaudSLLI)extent of
Priorit.y-based Ilro.-i’.lnptiollA
COllligul’atioll
whicht~SillliC..’-i that semi’nllliib<2r of Im,Vlouslynlad+..
a.~sigilnlent.s (’all b<’ relaxed (or disrupted) is ch’lined
by Ihe triph, <(;cl/Hrqu t,qt;dl-les,
(;¢’;gBa,il,
htts,
El:al.w~.m,,.i,
pt;,,,.,>.Tiffscon[igllral.i(lll inipielllents
a. fern< of I)re-ellil)lion,
basedon liiissioti
priority.
hi sc~uniing a candida.l.t,
wing’s Cal)a,’h,y profil,,.
(/cllBu,,q, lnt.+ considers capacity (:urrently allocal.ed
i.o lower prioril x., inissh)ns ~ available for assignnienl,
al,d geu<.,ral.es all,><’ation intervals Im+ecl on I.his assuinptinl~, li’vaIMinDis:.upti<:n
pronlotes allocation
hll.el’ValS that <lisrupt the fewest nlissiolls and Ihos+
with the lowest I)riorhv. This nlhiimizes the ca.s,:ading eftk, ct (since any mission that is prt~eillptt.d
by
a higher priority nlission is rectlrsively
re-scheduh,d
ushlg the same proc<’duro).
Gh’en the conlhhia.torial nuint)er +if allo,’at.ion iutervals iuld possible sets of bull<peel nlissiotls l.hat can
I)+" generated vi,’t a coulplete capacity prolilo scanniug procedure (O(e.r) where e is the ,:apa,:ity of the,
resource and f is the dur++tion of <’apacil.y profile
fragulents), our ,’urreul imph:nientation utilizes a [hleltr. hell risl ic sa n ipling st riltegy (O(cf)) COlillmtible
with EcalMi,tDi,,,.,,rtio,t.
Briefly. allocation intervals
;tre generated by single forward scan through Ill<’ resource’s cap~’ity prolile over the time interval where
capacity is required. At each tinie point encountered
duriug the scan. I, he set of pre-enlptalJle niissions is
,’oillputed. If uotl-elnpty, I.h,’ currellt alloeat.ion interval is extended by (I) <’OiUlml.ing the subset of preeliif~tal)h, nfissions of lowest priority an,I (2) sel<.cting
38
AIPS-2000
[.’igure :|: l~xplol’atiori of ,’on.-,tt~hil, r,.laxalion ol)l.i,,iis
using configural)lc, lssignllssion
l)ro(’t.thn’<..
I.he niissioit in this subst’t wil.h niaxilnal finish t.inte.
Coniposl/(; Relaxatlolis - (’onlp, m,,ili.s of ih,’ ;tl)ovl’
1)aso
<.’onfiguraliou~
c:-in :llso [.,e ,’,lllll)(.,,~t,tl
1,, ,l,,tin,’
conliguratiolls t)t" hssignXissioawh,,r,, illull il,h’ ,’OlisLrtfints are siinultan,,ously relaxed.
lsstgnX±ssion
(iii all)’ o1" t, he coitligliratinils ,h’s,u’ibedabovc’)canbe apl)lie,.l it) ally .,>t,h,cte, I s,,!.
Inissions vit~ the hssignl4issions llrocochll’< , given in
,, the
Figure ?. Within this CSP-slyh, search llrot-+,dtil’t
curro.nt, iilllllt.’nlenl,il,
lion nsesmissionI)riority aLid lal.,.~+r d,.Ih,ery dale as a heurislic basis for lilissi¢,ii s,,iecl.ion (i.e.. variable¢,’(lering). hipul niissions,’.it’,, sorl(.tl
on this basis +llit] hss±gnllission
is tlmll s<,Clll<,lllially
al)pli,-d to eacil to alloct+te requir,,d rescllii-<-es (ilSillg
wliicll<wer of the ahoy,, c,)nliguraliOus as I)e<.ii ,l+,siguated).
In caseswh<,roa giveui,d,,>.sion hasnufi.asible’, assigli-.,
III<’IIt.S
([.<’., AssignMission(M,
feasible)
is ;qq+lied
and returns no solutiori), all exploration or" pnssihl+, relaxation options can he condu,:l.ed through r,’l)eale, I al>pli,’ation of Assign~ission in di[l’<:r,:ut Collliguralions
(as depicted
ill l-’Jgure 3). Autolnalioti of I[iis i)rocess
i’<,quires a gh)ba] 0valuat[oli criterion siiilal.)le I’~,i" r,’laling opt.ions generaledai-i’oss di|lTi’enl iliv<,caliOli.~ of
hssignMission
(e.g., so<lie liie¢"lSlll’l,
tit" ov,.rall b,:nelit
lind cost). In the current inll)lonleuial.i(~u,
this pr,lcednre i,,, usediu w]i~t.-if nioclo to gel<oral,, all++rlull.[vv
optious, a, iid the iiser l’mr[’oi’iiis the ,,vahl;lliOli ai,l selecl.ion.
Mission Combination
A second ,:ore I’unclr, ioli provided hy the I~;~rrel AIIo,’ator
is mission couibinal.ion, a c~qmbilii, y quite analagous to
Iho mergiugof shnih~rjol)s in a job shol.)lt~ r,,di.tce selup
costs alid [iicrease resource;.wailabil[ty. Tit, hlillloiiieni;d.ion of this calmlfility exl.en,.Is the ha.sic ternlJl;ll.,, pl’¢sented in the previous imragraphs, Collpling tilt. t,xt,t:’tllion of a "planning" compo,,,nt that computes -ill,’,t’l~alive ,’,Jmposite Inissiou iti,eraries wit h a "’sche,:luling"
,-ompolieid, that verifies fe~,-sibility of a giv,’lL ,:ond.,ined
uiis,sion and geuerates a wirlg assigninenl. The sche(lul-
From:
AIPS 2000 Proceedings.
Copyright
© 2000, AAAI
(www.aaai.org).
All rights
binedreserved.
missions.
ing component
is the basic
AssignMission
already
de-
fined.
More generally,
mission eonlbination
can be
seen as a specific configuration of all integrated
planning and scheduling search template,
designated by the following triple of components:
¯ ,
¯1
t
<(,tnpossOomb~GtnFca,,t;ontb,l~J~,alMaxTripRed>
where
(/¢llFca.~t.’o,nb
is all instantiation of hssignMission.
In this case, the operator (;enpossc’o,nb represents tile
planning component. It will generate a set. containing
all possible composite missions involving a certain
mission M. For m~ypair of missions (:’t.ll,M,.,),
there
are two possible combinations: one in which mission
:’tlz flies after the end of mission M1"s last leg; and
one in which mission :’fit flies after mkssion M.,.
Possibl," combinations ,,inslsatisfy the conj,,nction
of constraints established by spatial and temporal
restrictions, plus constraints following from the three
exlernal parall|eters
specified by the user: mininml
pcrcenl.age reduction in total duration of the combined
mission, the n,aximmn layover time, and t.la~ maximum
distance between the end location of the first mission
and the start of the second. The output of this
operator ca_Aabe seen as Ihe set. of all notional combined
missions Mresulliug from ,’oncatenatio,,
of possible
cornbinat ion pal rs (,’tll, M.2).
Once all the possible
combinations
has been
gonerated, tim existem’e of a feasible wing as~igllnlent
wi]]__be delennin~,d via application of
hssignMission(3/,feasible),
where M represents
the mi,ssion generated by combiniugthe itinerary of two
mi.ssions in a possible pair (including an intermediate’
coutmeting flight leg if m:cessary). The allocation interval should be large enough to accommodate the duration of M. Although currently only feasible allocations
are considered, auy of the relaxed hn-~igaHissioa ins!antiations could be u,sed.
Once all feasible combinal ions have been determined.
the. candidate providing the largest overall reduction ill
airplane flying time is selected. Figure 4 summarizes
the procedure.
Ahhough CombineMission
allows only two missions
to be combined at a time, note that the combination
of larger numl~ers of missions can be accomplished by
recursively applying the algorithm to previously corn(.’.ombincMission(M, MaxLay. MaxI)ist, ReqRed)
(hulerate Possible (’.ombinatious:
For each Possible ComlfinationPair (M1..,’vt,.,):
Generate Mission M = :I[I+M1..,+M,.
For each M in Possible Combinations:
if hssignMission
(M, feasible)
Then add M
Apply EvalM,,~Tripma to Feasible
End
Figure 4: Mission Combination Procedure
Linking Tanker Missions
The linking of tauker missions to airlift missions is similar t.o mission combination. For a mission requiring
air refueling, the algorithm first tries to find an existing tanker mission that is already scheduled Io be in
the vicinity of the requested refileling track in the time
period requested. In the event that ntultiple tanker
missions art, found, the mission that minimizes the perturbation in both missions will be selected. Failing to
find an existing mission, a new I.anker mission will be
created and linked to the requesting mission if there is
available tanker capacity.
The search template for air refueling can Ix, expressed
as <(.~’er$Po.~s Refuel ,(’;~tFeasRefu:.l
..Et’alMin Pert >. Still-
ilar I.o the Mission Combination. the operator
(~en.l, ossRelucl corresponds to I he plamiing component.
It will search for the set of currently planned refiteli,g
missions that could possibly service a ,’ertain mission
M. The linkage of tankers will typically require additional adjustments of start m,d end times of both missions to guarantee the synchronization of tile legs and
resources involved in the air refilcling activity. This is
achieved by establishing temporal constraints betwee,
the activities representing fuel supply and reception.
’File time bounds of both missions involved are l,h~n
updated by propagating these constraints to the entire
mission itinerary.
In contrast to Mission Combination,the "’,-.cheduling"
COml)onent
here. (re’lll.’easRt./~tcl,
iS composed
of twoinstances of hssignMission,
to m;tke airlift aml ta.nk,.r
wing a.ssignments respectively. Ouce the set of possible refueling rnissioos has been identified, and the time
bounds of the missions have been adjusted accordingly.
Genl,’easReluetwill generate the set of time intervai,~ for
which both the airlift, and tanker resource are available
simultaneously during the entire duratkm of the refileling. If there are consistent wing assignutents for I)oth
missions, the pair will be markedas feasible.
After all feasible pairs havc. beenidentified, tilt. evaluation function is used to select the pair thai, n,inimiz~-s
perturbation in the missk)ns’ itine.raries. If no feasible
pair is found, and there is enough tanker capacity. ;,
newreflleling mission servi,g the airlift mission is created. 1 Figure 5 summarizes the air refueling linkage
proccdure.
Utilizing
the Core Components
The three
core procedures
discussed
above.
AssignMission~
CombineMission
and LinkTankers.
are composed in various ways to provide user-level
functionality. I"ach can be inw)ked individually on
I With regard to current intended use, the creation of a
newmission is equivalent to reporting failm’e. The creation
of a dunuuymissiou is the mechanismthat can be used to
notify the plmmerthat there is a mission that carmot be
refueled giving the existing set of missions.
I.i,kTank,,rs(M)
I.hereserved.
AMC
(’.orpor~d.e
From:
AIPS 2000 Proceedings. Copyright © 2000, AAAI (www.aaai.org). All rights
Generate Possible I~.eI’ueling Links:
Find possible pairs (M.R)
Adjust ~rlltt l)roi~agate, lime boundsfor pair (M.R)
For ,:ach Possible Refueling pair (M.R):
If AssignMission(M,
leasible)
and AssignMiss
ion (R, feasible)
Thenadd p~ir(M. R) to Feasible.
If l"ea.,~il)le is not empt.y
Theutapply1.5"FttlMi,li.,ert
tO l"~’asiblt,
Else generate new refl,,ling mission
End
Figur,, 5: Air Roflwling Linkage
giv,,n soh,,:tcd mission: this is thetint,st
granularily
modeof inLeraction with the system. Mort, typically.
howew’r,user a,’tion is tt~ken relative to sea,’ selected
set. of input missions. The AssignMissions
l.,rore~lur- (Figure ??) provides a bxsis for simuhalleo,sly
tdlo,’at.ing resources Io $olne sel of tnission.q under
;L given set of constraint relaxation assumptions.
Ana.lagous ,’OltLposite procedures are simihwly defined
and provided for CombineMission, to allow overall
,’,)ll,pt’es,.,iol, ,.,f i’(.SOllrCO
usageow,rSOlll(: interval,
for LinkTankers,t o allow sinlultaneous I, reat Jnent of a
set of unsatisfied air r~.fiwling requests.
At. i~resenl., taw configuration of ftm,’tional eapabilit.i,,s
for mission allocation, eonlbinalion and linkage for refueling is under user-control.
Typically,
Ass ignMissions is used in f’¢msible assignment modeto
conslr,t,’t a base allocation, interh’aving the us,: of an
AssignTankersromposite as needed t.o reconcile refitcling re, luirement.s. CombineMission
(or its contposite
(.’o]lnlerpart CombineMissions)
are th,,n applied in ,’on.i,,lt,:tion wi!]l va, rious relaxt,d forms of AssignMission
toit,:ral iv+,ly inLprov,,,’eSOllrrt~ us,+g,, and ac,’olmn,>d~u.e
addiLional, lower priority tnissions. Oil-’ a.rea of curry,hi
workis l.h,’ design of il.,,rative inlprov<,nlentscar<’h pro,’edtm.-.s for automatingt.his pro<’ess.
Implementation
and Status
The scheduling <,ngine of the Barrel Allocator is imple,l,,nl.od
in Allegro CommonLisp 5.01. Tile ilser
interface is in Java 1.2. The curl(at version rtlns on
WindowsNT plal.fornls.
The yore mission sch<’duling
procedure is quite efficient a 2 week i,tervml of ntissions extrax:ted from the current. Corporate datt~ base
(approxitn;ttely 1001) missions, 5000flights) is scheduled
from scratch in less than 20 seconds on a Pent, lure II
400MI Iz. Increment al planning and scheduling all.ions
are exe,:ute¢l in real-lime.
With support, from Logicon Corporation, the principal developer of AMC’sConsolidated Air Mobility
I’lanning System (CAMPS),the NTversion of the Liarrel Allocator has been integrated to communicatetransparently with other CAMPScomponetlts and to support AMC’soverall business process. Connections to
40 AIPS-2000
I);tta Bast’ and to oth,:r (’AM
tools areaccomplisl-’d thrtmg, h a COM
int~rfa.,’c.. The
hlarr,,l Allo<’a.l.or
,’,)lllll, OllellliS illll)l,"nt~’nl,’d:t., a (’OM
s<.’rvt+l’ that ran I)e <:~dl,.d froma.lty clio,t in (’A.MPS.
It a.lso Illakes Its, + of sew~ralCOM
s~.rv,.l’S :~vaila.l,h. in
CA).II’S.
Tl,e AllocatorI,,a, ls i’o~our,:t, availal,ilitv an, I ittissio,
descriptiondat ;~ fromtile AM(’ ( "orporat,. I’):41 a has,,
ing one such (!OMsorv,-r st.rver pr(,vided by (’.AMPS.
Thisserv,.r makesIhe querie.~ to l.he dat;dmsoat,.[ t ran.sh~tes missiol~ and r<,sour(’,, it~formationil,l.o a t,)rnlal.
the Allocator can list,. Theuser sp,,,’ilios :t tim~’ int,,rval. and a certain s,.I of air ba.,~,.s an, I air,’rafl lyp,,s
of intr’l’OSL. Thesyslemwill t.h,’n, l.hl’t,u~h th,’ (’().XI
serv,,r, query1.he dat.abase for n,,ntinal r~.s,,Ire,, availabilityl,,vels (i.,:., tmntb,.rsofcoltt.r;Lel.,t’~.ll,’,.,I :llld I,’~sst.ssedaircraft.) and;dl ,,fissions I’,’qilirillg I lit’ it.",,’ ,’~1’
I.hose reSollr(’es duringt l,. Sl,.,’iliod titlto i,t,.’l’V;,I.
"1"ostq~porl,colttinuousdaily Ol)er;d.iOllS.1.h,. ~nissi,,ns
retrievedfromflit, tl:~tal.m..4eare ulark,.dI~al’l.il i,,t,.tl ilfl.~.,
threel)ossiblest at,e,-,:
¯ .\’~ w missiom~:missionst,ewly rt’,.ated or im,,lifivd by
the plannerandl)eltcling a.i)pr,,val I,y th,, I~arl’,.I.
¯ ..Ipi,ror¢d Mis.~itm.¢:
n~i.,,si,,ns t.hat, the. Ilarr,.I all,I Ih,,
phumerha.w’I’e;wh,,d cOns¢.ll~llS:,.l,I [;,r wl,i,’h r,.stmr,’-s hay,. I.,en all,.,.’aled.
¯ RevisedMi.s.~io,s: missi<.,ns t.hat Ihe It:,l’rt.I has alh,r:tl.t~d resourres for tutd,.r rolaxe, l assulnl~li,,l,S ’ but
tile planner ha.s not yet ,:oneUl’l’ed.
Silly ttpprot:td atld re’vistd tuissionsare Ill,’ ini..,si,:,,s
LhcB;trl’el has prt~viouslyalloc;~re,.I. Ih,,y will b,.. ;ultomari,’ally assigned, during load lira,’. Io Ill,. wing
and lille interval SlW,’ifiod in Ihe i’equt.st. "1"1,, r,,._
l.ual resource;wailability level is ol,tain,.,l byt’,.,l,t,’iI,:g
l.he I~ominalavailability by lh,. at~tOll,t ~,f Cal).,’il.v r,’served by already allo,’atcd n~issio,s. Th,. syst,,nl will
flag a conllirl if chang,.sin nominali’t.SOlll’t.’c :,vaihd,ility causea wingto be. unavailal~hrI’~,r alr,.atly a..,sigl,’d
missions. Those problematir missions will 1,, nmrk,’d
as uuassiynable and will rr, quir,, senv: user guided ;wtion to 1)(, reintroducedil,t.o the current schedld(,. Once.
the barrel has finished w(,rking with a s,.= ¢,t" luissi,~ns.
s/he can rotumit the derisions bark t.o !lie database.
Missions that h;we had Lheir flying I.ilues or assigned
rt.sotu’r,,s rhangodbyth,.. lmrl’ol, will I:,c. In;u’k,.d as revised, pentling apl~roval nr further rhang(, I,v !h,. i’,,levent planning ofli,’e. Subseqltently ,’alwt.llt.d missions
missionsthaL the plann,,r has decid,,d will tat, 1,mgr.r b,.
supported- will be rentov,’d froml.ht, srh,,,lul,,. I"fising
I I,’ opportunil, y l o reassossother, pr,.vi,.ms ,"tmSt.l’aint
rela.xal.ion d~cisions.
Themostr,:t.’ent v(’rsioJl of the Ilarrel .’%llo,’at.,~r was
transferred to AM(’.in tnid October1!)!)9 for alpha testing. It is oxpt,cte, I Lo be released as an operat.ional
contponenl of CAMPS 2.0 by March 2001).
Conclusions
From: AIPS 2000 Proceedings.
Copyright © 2000, AAAI (www.aaai.org). All core
rights techniques
reserved. are
In this paper we have described Barrel Allocator, a
mixed’initiative
system for day-to-day managementof
airlift and tazlker resources. The system has been designed to provide a range of mission scheduling and
planning capabilities, including incremental generation
of feasiblc resource assignments for pending missions,
generation of alternative allocation options in casc of
resource contention, identification of opportunities for
marc efficicnt aircraft utilization though mission combination, and generation and synchronization of tanker
missions to meet air refueling requirements. Each of
these capabilities can be utilized with different degrees
of user control ow~r the decision-making process, ranging from user-controlled option generation to fully automated scheduling a~ld planning processes.
The Barrel Allocator implementation derives from
the architectural principles, scheduling ontology and associated class library of the Ozone scheduling framework. which consolidales the results of application
building experiences in a number of similar problem
domains. Although, any new problem domain brings
uniquc requirements and constraints that make it difficult to use pre-existing solutions "out of the box",
this starting point has nonetheless substantially accelerated our efforts to develop and transition the Barrel
Allocator application. The core procedure for mission
allocation, for example, was initially prototyped as a
fairly direct instantiation of a search template previously developed in another transportation scheduling
context, and allowed rapid development of initial mission combination aud linkage capabilities.
Later on.
the architectural and configuration flexibility of the tmderlying frameworkhas allowed us to efficiently refine
this functionality and to quickly respond to requirernent changes re,~ulting from increased user exposure
and evolving AMCbusiness processes and policies.
domains(Becket 1998) have flexibility
C.oncerning the effort involved in this transition, the
greatest challenge has been, and remains, to create a
bridge between the current culture and 1)usiness processes at AM(’., and howsuch processes and culture can
and should evolve as the Barrel Allocator becomesfilly
operational. To gain nser acceptance it has been necessary on one hand to demonstrate the ability to support
current processes. On the other hand, it ha,s been necessary to demonstrate alternative business processes that
better exploit the pol~ential of the Barrel Allocator to
improve decision-making within AMC.
As the Barrel Allocator transitions into operation at
AMC,several extensions to system flmctionality are
currently planned. One near-term extension will ex,tend the system’s resource allocation mechanismto additionally consider air crew capacity and availability
constraints (currently we enforce only constraints on
crew duty day and rest requirements). A second direction of future work is to expand the system’s "reactive" capabilities and support shorter-term response
to exceptional execution events. Muchof the system’s
directly applicable to this rcal-t imc,
execution management process. A third direction of
planned extensions concerns evolution of the inl.era,’tire option generation process. One particular interest
is in developing better techniques for visualizing mid
comparing the impact of change.
Acknowledgements
The current Barrel Allo(:ator system represent.s the ,’umulative efforts of several individuals. Dirk Lem,,,:rmann. Gary Pelton. David Itildum. Mark Shieh. and
Seppo Torma have all made substanlial conttihuli~,ns
to the implementation. Mark Burstein hms been instrumental in the definition of protocols for integral ion with
the AMCCorporate data base. and in steering the overall integration elfort. Brian Gloyer. Lindy Resner and
others at I,ogicon have provided great support in interfacing with various CAMPS
system components, and in
understal,ding user requiretnents. This work has beer,
funded in part by the Depart naent of Defense Advam’ed
R,esearch Projects Agoucyand the US Air Force Ron]~.
Research Laboratory under contracts 1;30602-97-2-0227
and F30602-96-D-0058 and by the CMI" Robotics Institute.
References
C. Barnhart, N.L. Boland, L.W. Clarke, E.L. Johnson.
G.L. Nemhauser.and R.G. Shenoi. Flight string mo,.lels for aircraft Ileeting an,I roul.ing. Traa.sport(tlw,
,qcieuce, 32(3):208-220, August 1998.
M.A.Becket. Ret.’onfigurableA rchih.ct ut~. s J’or Mi,redInitiative Planning and Scheduling. Phi) the.’,is. Graduate School of Industrial Administration and The
Robotics Institute, Carnegie Mellonl;niversity. Pittsburgh, PA..luly 1998.
C!larke L.W., E.L. Hane C.A., Johnson, and (.l.[,.
Nemhauser. Maintenance and crew co,.,s.iderations
in
fleet assignment. Tlvmsportation .g(’iencc, 30:2-19-261),
1996.
R.A. Rushmeier and S.A. Knotogiorgis. Advam’es in
the optimization of airline fleet assignmem. 7i’a,.sportation Science. 31(2):159 1(i9, May19!)7.
S.F. Smith and M.A. Becket. An ontology for constructing scheduling systems. In PuJcetding.~ of the
AAAI Spring Symlu~sium on O,lological Engim criag,
pages 120-129. Pale Alto. CA, April 1997.
S.F. Smith. O. Lassila.
and M.A. Be,:ker. Configurable, rnixed-initiatiw, systems for planning and
scheduling. In A. T’ate. editor, Advanced Planning
Technology. AAAIPress, Menlo Park, 1996.